# Kolibri-1 ![Organization](https://img.shields.io/badge/Org-Aleph%20Alpha-000000) ![License](https://img.shields.io/badge/License-Apache%202.0-2da44e) ![Params](https://img.shields.io/badge/Params-78B%20(3.5B%20active)-8957e5) ![Context](https://img.shields.io/badge/Context-1M%20tokens-blue) ![Released](https://img.shields.io/badge/Released-Oct%202026-555) > **Quick answer:** Kolibri-1 is Aleph Alpha's open-weight mixture-of-experts reasoning model with 78B total and 3.46B active parameters, released October 3, 2026 under Apache 2.0. It is trained from scratch on 20T tokens with a German/English focus, supports reasoning effort levels and tool calling, and a 262K-token native context validated up to 1M tokens. Aleph Alpha reports 84.3 on GPQA Diamond, 85.9 on LiveCodeBench v6 and 66.4 on SWE-bench Verified. ## At a Glance **Where it leads:** Strong math and coding for its active size (AIME 2026 96.0, LiveCodeBench v6 85.9) and German-language reasoning, with FP8 weights that fit on a single H200 or B200. **Where it lags:** Behind Qwen3.8-27B on most of Aleph Alpha's own comparison table (e.g. Terminal-Bench 2.1 27.7 vs 76.8, Humanity's Last Exam 21.5 vs 35.6), and weak on AA-Omniscience factual recall (accuracy 14.8). **Best for:** Sovereign German/English assistants, retrieval-augmented generation and agentic tool calling where inference cost per token matters. ## What Kolibri-1 Is Kolibri-1 is a 50-layer MoE transformer with 384 experts per layer (top-6 routed plus one shared) and a 4:1 sliding-window-to-full-attention layout, trained with Muon on 768 NVIDIA B200 GPUs (about 21 days of pre-training, 6.4e23 FLOPs). Post-training combined supervised fine-tuning with asynchronous RL across reasoning, tool-calling, software-engineering, terminal and retrieval environments, including German-language environments with a language-consistency reward. It exposes explicit reasoning effort (`none`, `low`, `medium`, `high`) through the chat template, ships with a custom German-aware tokenizer (UniBPE, 128K vocabulary), and is released as FP8 weights with a BF16 variant. Aleph Alpha is a signatory of the EU GPAI Code of Practice. ## Specifications | Field | Value | |---|---| | Organization | Aleph Alpha | | Hugging Face | [Aleph-Alpha/Kolibri-1](https://huggingface.co/Aleph-Alpha/Kolibri-1) | | Architecture | Mixture-of-experts, 78B total / 3.46B active | | Languages | German, English | | Context length | 262,144 native; validated to 1,048,576 | | License | Apache 2.0 | | Release date | 2026-10-03 | | Knowledge cutoff | 2026-06-18 | | Reasoning mode | Yes (none / low / medium / high) | ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [GPQA Diamond](/benchmarks/rein-et-al/gpqa-diamond) | 84.3 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | [Humanity's Last Exam](/benchmarks/center-for-ai-safety-scale-ai/humanitys-last-exam) | 21.5 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | [MMLU-Pro](/benchmarks/tiger-ai-lab-university-of-waterloo/mmlu-pro) (CoT) | 80.0 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | [AIME 2026](/benchmarks/mathematical-association-of-america/aime-2026) | 96.0 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | [AIME 2025](/benchmarks/mathematical-association-of-america/aime-2025) | 96.9 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | [LiveCodeBench v6](/benchmarks/livecodebench/livecodebench-v6) | 85.9 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | HumanEval+ | 92.7 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | [SWE-bench Verified](/benchmarks/princeton-nlp-openai/swe-bench-verified) | 66.4 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | [Terminal-Bench 2.1](/benchmarks/tbench/terminalbench-21) | 27.7 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | [IFBench](/benchmarks/allen-institute-for-ai/ifbench) (loose-prompt) | 78.1 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | [BFCL v4](/benchmarks/patil-et-al-uc-berkeley/bfcl-v4) (overall) | 61.4 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | [AA-LCR](/benchmarks/artificial-analysis/aa-lcr) | 68.3 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | [Tau2-Bench Telecom](/benchmarks/sierra-research/tau2-telecom) | 94.7 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | [Tau2-Bench Retail](/benchmarks/sierra-research/tau2-retail) | 69.9 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | | [Tau2-Bench Airline](/benchmarks/sierra-research/tau2-airline) | 76.7 | [Model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) | 2026-10 | *Self-reported by Aleph Alpha; not independent Benchgen measurements. The card also reports BrowseComp 29.4, Tau3-Bench Banking 38.1, FRAMES, SealQA, MuSiQue, LongBench Pro, RULER, HELMET and German-language variants of several benchmarks; these are not added because the setup differs from the Benchgen pages (e.g. Qwen3.5-35B-A3B's BrowseComp is 36.5 on this card vs 61 on Benchgen) or no matching page exists.* ## Frequently Asked Questions

What is Kolibri-1?

Kolibri-1 is Aleph Alpha's open-weight 78B mixture-of-experts reasoning model (3.46B active parameters) for German and English, released October 3, 2026 under Apache 2.0.

What is Kolibri-1's context length?

262,144 tokens natively, with quality and serving efficiency validated up to 1,048,576 tokens.

Can I run Kolibri-1 locally?

Yes. FP8 weights need about 78 GB of memory, for example 2x H100, one H200, or one B200, using the aleph-alpha-inference vLLM plugin.

Kolibri-1 — Evaluations

Evaluations